Business Understanding

Divvy, Chicago’s bike-sharing service, faces fluctuating demand for rides across days, weeks, and seasons. These fluctuations impact bike availability, station balancing, and operational efficiency. Accurately forecasting ride demand is essential for ensuring bikes and docks are available when and where riders need them. The business objective is to develop a time series forecasting model that predicts short-term and long-term ride demand, helping Divvy optimize resource allocation, reduce service disruptions, and improve customer satisfaction.

Problem statement

Divvy, Chicago’s bike-sharing system, faces dynamic demand patterns that vary by season, day of the week, and rider type. These fluctuations often lead to bike shortages at high-demand stations and surpluses at others, reducing customer satisfaction and increasing operational costs. To address this, Divvy requires accurate forecasts of daily ride demand. Predicting demand trends will enable better resource allocation, efficient bike redistribution, and targeted marketing campaigns, ensuring improved rider experience and sustainable system operations.

Data Understanding

The dataset consists of Divvy trip data from January 2024 to the present, including ride start/end times, station information, user type (member vs. casual), and trip durations. Since rides are timestamped, the dataset supports the creation of aggregated time series (e.g., daily, weekly, or monthly ride counts). Additional contextual data such as weather conditions, holidays, and day-of-week effects can be integrated to better capture external influences on demand.

Data Preparation

Data will be aggregated into time series at different granularities:

  • System-wide ride counts per day/week
  • Station-level ride counts
  • Ride counts segmented by user type (casual vs. member)

Missing values, anomalies (e.g., extremely short or long rides), and seasonal events will be identified and handled. Features such as lagged variables, rolling averages, and holiday/weekend indicators will be engineered to strengthen predictive modeling.

Modeling

Time series forecasting methods such as ARIMA/SARIMA, Exponential Smoothing (ETS), and Prophet will be applied to capture trend, seasonality, and holiday effects. Advanced models such as LSTM/GRU recurrent neural networks may be explored for capturing non-linear temporal dependencies. Models will be trained and validated using rolling-window or walk-forward validation to mimic real-world forecasting scenarios.

Evaluation

Models will be evaluated on accuracy metrics including RMSE (Root Mean Squared Error), MAE (Mean Absolute Error), and MAPE (Mean Absolute Percentage Error). Forecast interpretability (e.g., identifying seasonal effects, day-of-week patterns) will also be considered to ensure insights are actionable for Divvy’s operations and marketing teams.

Deployment

The final forecasting model will be designed to generate regular demand forecasts (daily or weekly). Forecast outputs can be integrated into Divvy’s decision-making process for:

  • Station rebalancing and bike redistribution planning
  • Staff scheduling and resource allocation
  • Targeted promotions/marketing campaigns (e.g., weekends, holidays)